Jer Clarke3d ago
Dare Obasanjo

While his framing is a bit cynical, I share similar irritation to Ed Zitron on AI companies pushing loops.

Loops where Claude or OpenClaw keep trying to solve a task until successful are essentially a token burning workaround to deal with the fact they don't get the answer correctly on first try.

As a software person, my inclination would be to fix the product and not have customers run infinite retries until the tool gives the right output where each retry costs money.

Jun 25 at 8:19pmWeb
Show threadFrobnitz Coil Oil5d ago
@carnage4life The only way to win is not to play.
Show threadevana5d ago
@carnage4life this is why you're not a paper billionaire.
Show threadGabriel Pettier5d ago
@carnage4life but the "flacky" lib/decorator has been insanely popular for a reason, mindlessly throwing more resources at a problem rather than fixing fundamental issues is very convenient sometimes, and by that i mean almost all of the time, until you get a monster down the line that no one understands, but that's for the next guy by the time you have jumped 3 jobs.
Show threadklausfiend5d ago

@carnage4life An algorithm that doesn't solve the problem is normally referred to as "a failure."

Asking Claude to iterate on its mistakes is not a "workflow", it's fucking irritating.

Show threadMikael Lundin 🍀🥦♻️5d ago
@carnage4life they don’t really have an incentive to make it better when they charge per token.
They want to give better result compared to their competitors while spending as much tokens as possible. That’s the whole business model.
Show threadmodulux5d ago
@carnage4life It would be better to get the correct solution on the same go, but it's unclear to me that anyone knows how to make that happen. There are all kinds of problems that don't have clear closed-formed solutions where iterating and occasionally backtracking is necessary. It's fine to say "we could just solve it right instead of wrong," but not at all straightforward to actually do that.
Show threadPeter Bloem5d ago

@carnage4life The result of a long reasoning trace can be distilled back into the model next time to shorten the reasoning trace, or even have the model one-shot the solution.

This changes the economics of the argument. The customers pay for the loops which is the most expensive part of the training. Anthropic pays for the distillation (and subsidizes thereasoning), and the next generation solves things more quickly.